Systems often provide content based upon a user's past behavior. Whether the system is utilized for selling movies, music, books, etc., the system typically stores a record of the user's activity, e.g., an online purchase, at a particular point in time. During a later interaction, the system may provide content, e.g., recommend a product and/or service, to the user that is similar to the user's prior purchase. Providing content that is actually useful and timely to the user, however, tends to be more complex than what may be accurately represented by a single event context learned from an isolated instance of user activity.